portfolio

Tests that file their own bugs, and agents that fix them

On LoadBetter's SpotBot, a loop runs every two hours: end to end tests against staging, screenshots of every screen reviewed by Claude, GitHub issues that open and close themselves, and coding agents that pick those issues up and open fix PRs. The same loop is now being rolled out on Refyner. The only step a human keeps is the merge.
AI Agents
Infrastructure
Tools
Claude
Python
TypeScript

The problem with fast code

AI coding agents changed the bottleneck. Writing a feature is no longer the slow part; knowing that everything around it still works is. The bugs that hurt in production live between systems: an email that never arrives, a call that negotiates the wrong rate, a screen that renders NaN, a keyboard that covers the send button on one specific phone. A green PR check says nothing about those.

The loop

  1. Every two hours, a full end to end run hits staging: API, web, calls and mobile.
  2. A screen tour captures every page on desktop and phone, and Claude reviews the screenshots like a human tester would.
  3. Failures become GitHub issues automatically, and close themselves when the test passes again.
  4. Coding agents pick up the issues, write the fix plus a new test, and open a PR.
  5. A human reviews and merges, staging redeploys and the same test verifies the fix.

Results in the first week

  • ~4,200 automated tests on SpotBot, from unit to end to end
  • 68 GitHub issues filed by the machines in the first week
  • 52 agent fix PRs merged after review in the first week
  • A full run every 2 hours on staging